The starting-keeper rating is informative beyond what the team-level Dixon-Coles / Hierarchical-Poisson fits capture. Applying it as a predict-time defensive multiplier on the opponent's expected goal rate — λ_a_adj = λ_a × exp(−α · gk_rating_h_centred) and symmetric on the away side — should sharpen the ensemble's per-fixture probabilities without disturbing the team-level fits themselves.
The change was originally landed against a 2-walk holdout. The [handoff doc](../private/2026-05-24-ds-h…